Page 481 - ISC PROCEEDINGS 21.4

not only as a generative output, but also as a governed and reusable digital asset that can
support supervision, experimentation, and cross-institutional research.
Overall, this methodology allows the paper to move from a descriptive survey of
Financial GAN architectures to a more critical synthesis of their economic role and
deployment conditions. By combining technical classification with infrastructure-oriented
evaluation, the paper identifies both the strengths of current Financial GAN approaches
and the remaining gaps in privacy, governance, and real-world applicability. This approach
also supports the later sections of the paper, where the discussion of benchmarks, policy
recommendations, and deployment requirements is framed around the practical needs of
financial institutions and regulators.
Although recent reviews have made important contributions to the literature, they
differ from the present study in scope and emphasis. Wilson and Azmani’s systematic
review focuses specifically on GANs for financial data generation and market modeling
between 2019 and 2024, synthesizing 30 papers across four databases and highlighting
common challenges such as mode collapse, training instability, and regulatory
concerns(Wilson and Azmani,2026). Lee et al. adopt a broader generative-AI perspective
across finance, using BERTopic to map major themes in the field, with particular attention
to finance-specific LLMs, GAN-based synthetic data, and the need for regulatory guidance.
(Lee et al ,2024) .Eckerli and Osterrieder provide an earlier overview of GANs in finance
and include a proof-of-concept evaluation of three architectures on financial time series,
showing that GANs can reproduce useful statistical properties but still face practical
limitations.(Eckerli and Osterrieder, 2021) In contrast, the present review contributes a
more infrastructure-oriented and theory-driven perspective: it links Financial GANs to
information asymmetry, data governance, and digital infrastructure theories, classifies
model families by the stylized facts they capture, and evaluates them using not only
statistical fidelity but also privacy, auditability, downstream utility, and deployment
readiness. This makes the present study stronger in connecting the technical literature to
the economic and institutional role of synthetic financial data in the digital economy.
4. Financial GANs: technical overview
4.1. Autoregressive & recurrent-based generators
Autoregressive and recurrent-based generators (RNNs, GRUs (Chung, J. et al., 2014),
LSTMs (Hochreiter, S. & Schmidhuber, J., 1997), TCNs (Lea, C. et al., 2016)) embed time
dependency directly into both generator and discriminator so they naturally reproduce
short-to-medium range autocorrelation and some volatility clustering; variants such as
QuantGAN (Wiese, M. et al., 2020) and TCN-enhanced generators (Radford, A. et al., 2016)
extend this idea to handle transaction imbalances and regime structure, making them
well-suited for modeling momentum, mean reversion, and other local temporal behaviors
in market series.
4.2. Conditional and hybrid time-series
Conditional and hybrid approaches combine conditional/tabular sampling with
temporal decoders or explicit covariate conditioning (market regimes, macro indicators,
customer attributes), which helps the generator respect event-driven shifts and
heterogeneous subpopulations; methods in this group (e.g., CTGAN-style conditional
samplers (Xu, L. et al., 2019) and TRGAN-type models (Zakharov, K. et al., 2023) with
conditional vectors) are especially useful for regime-dependent volatility, scenario
generation, and producing high-fidelity slices of the distribution conditioned on
explanatory variables.

480